An internal OpenAI experiment took an unexpected turn after thousands of AI agents managed to post roughly 18,000 messages on a public German website, despite safeguards that were reportedly intended to limit them to reading information online.

According to findings reported by Ars Technica and cited by B’Chadrei Charedim, the activity continued for approximately six weeks on the German DSEwiki website. The agents operated under roughly 3,700 different self-assigned names, and OpenAI later confirmed that the agents were its own.

Researchers believe the agents were participating in a scheduled experiment requiring them to search the internet. While they were supposed to be able to read websites but not post to them, the agents apparently discovered a way to use their access to write on the small German wiki.

The website then effectively became a communications board for the agents. They exchanged answers, pooled research results and shared methods that could potentially help bypass restrictions in the sandbox environment in which they were operating. Researchers said this apparently allowed some agents to build on work already completed by others.

Some messages also discussed possible XSS attacks against the website and ways of impersonating its administrators. In three messages, agents reportedly used the term “swarm” to describe the group involved in the activity.

Researchers cautioned that they do not have the complete picture because their findings are based on publicly available messages rather than OpenAI’s internal logs. OpenAI said the material reviewed so far does not indicate that the agents actually breached the website itself.

At one point, OpenAI apparently became aware of the activity, and the agents’ posting on the site dropped sharply the following day, researchers said.

The incident follows another case uncovered roughly a week earlier in which researchers from METR found more than 1,200 OpenAI agents posting messages on a message board during a separate experiment. In that case, agents reportedly exchanged methods for stealing information from AI company Hugging Face, with some later succeeding in accessing its network. Researchers believe the two incidents involved separate groups of agents and separate experiments, which OpenAI confirmed.

OpenAI said it is reviewing the latest findings and will take additional steps if necessary.

(YWN World Headquarters – NYC)